cisco and ecole polytechnique · automatizing vulnerability research to better face new software...

Automatizing vulnerability researchto better face new software security challenges

CEDRICTESSIERINSTRUMENTATIONTEAMLEADER/[email protected]

Innova&on&ResearchSymposiumCiscoandEcolePolytechnique

8-9April2018

Innovation & Research Symposium 2018 - Cisco and Ecole Polytechnique

Software Security

• Data security depends on secure software

• Software contains bugs

• Some bugs are vulnerabilities

• software intended behaviour can be abused

�2


Vulnerabilities

• Unknown vulnerabilities will be discovered

• so-called 0 days

• A lot of them independently by several peoples

• contrary to popular opinion

• 0 days will be exploited in the wild

• NSA or CIA leaks

• Ransomware (WannaCry ETERNALBLUE)

�3


Vulnerability research

• motive (why)

• attack surface (where)

• knowledge (how)

• first move (when)

�4

Vulnerability research cannot be reserved to the bad guys…… as it will give them the advantage


Offensive Security

�5

• Deep complementarity

• Counterbalance bad guys advantages

• Increase the cost of attacks

• Knowledge is power“Ignorance has taken over Yo, we gotta take the power back!”

Rage Against the Machine

From a defensive only security paradigm…

…to both defensive AND offensive


Auditing Software

�6

Auditing software and finding vulnerabilities is crucial

“Who looks outside, knows nothing; who looks inside,glimpses the incredible waiting to be known.”

Carl Snow


Platforms Diversity

• Huge diversity of platforms

• toward the end of Wintel (Windows + Intel x86) era

• ARM's dominance on mobile markets

• MIPS, PowerPC, [your 90s architecture] still kicking

�7


Software Complexity

• Increasing complexity of the applications

• multi-megabyte software libraries are common

• web browsers are more like small operating systems

• Closed source binaries

• very common in the industry

• require reverse engineering

• but fewer eyes often means more bugs…

�8


Increased Difficulty

• Overall improvements over the past years

• more and more mitigations and compiler enhancements

• better development cycles (continuous bugs hunt)

• Finding exploitable bugs is more difficult

• low-hanging fruits less and less common

• yes, it’s bad news (think as a James Bond villain)

�9


Costs

• Bad guys have resources (sometimes much more than you think)

• criminal organizations

• state-sponsored groups

• military and secret services

• Good guys have limited resources (sometimes even less than you think)

• time (money)

• workforce

�10


Finding vulnerabilities

• Never-ending quest (growing code base)

• Renewed challenge (increasing difficulty)

• Competitive field (inflating investment)

�11

Innovation is mandatory

New tools and strategies are needed


Binary Analysis

• Dedicated tools

• disassembler

• debugger

• Specific techniques

• static analysis

• dynamic instrumentation

�12


DBI

• Observe any state of a program anytime during runtime

• Automate the data collection and processing

�13

“Transformation of a program into its own measurement tool”


QBDI

�14

QuarkslaB Dynamic binary Instrumentation

• Open-source

• Cross-platform

• macOS, Windows, Linux, Android and iOS

• Cross-architecture

• x86_64, ARM (more to come)

• Modular design (Unix philosophy)

Give it a try! https://qbdi.quarkslab.com/


Modularity

�15

• Only provides what is essential

• Don’t force users to do thing in your way

• Easy integration everywhere


Integration

�16


Fuzzing

• Fuzz testing software (aka fuzzing)

• injects randomized or mutated inputs

• provides a way to find bugs

• Completely automated

• input generation

• software execution

• crash (pre)analysis (or triage)

�17


AFL

�18

• State-of-the-art fuzzer

• a reference in industry

• impressive trophies (openssl, openssh, …)

• Open-source© Michał Zalewski


AFL

Guided Fuzzing

�19

INPUT BINARY PATH

Generate Execute Record

CRASH

Feed back

MUTATED INPUT


Smart Fuzzer

• Hybrid approach

• various brute force strategies (input mutation)

• genetic algorithm (input selection)

• Focus on inputs that produced new path

• Maximise code coverage (better results)

• Minimise search space (less time)

�20

aims at better efficiency


AFL Limitations

• Pros:

• Fast (scale for thousand executions per second)

• Efficient (find bugs in real-world applications)

• Cons:

• Targets sources are required

• Portability

�21


AFL/QBDI

• Targets closed source binaries

• Allows runtime optimizations (search space reduction)

• Reverse engineering needed (no sources)

• often minimal but mandatory when targeting internals

�22

AFL with QBDI as the instrumentation engine


Fuzzing Binaries

�23


Symbolic Execution

�24

• Analyzes software without running it

• Uses symbolic values instead of inputs (abstract interpretation)

• Represents computations as expressions

bvadd

bv

1 8

bvxor

bv bv

10 8 20 8

mov al, 1 mov cl, 10 mov dl, 20 xor cl, dl add al, cl


Triton

• Cross-platform (macOS, Windows, Linux)

• Dynamic Symbolic Execution (DSE) engine

• Integrated constraints solver interface

�25

Open-source dynamic binary analysis framework


Constraints Solving

�26

• Taking a path or not depends on conditions

• Conditions create path constraints

• Symbolic expressions can represent constraints

• Constraints can be solved symbolically (SAT solvers)

y = input[0]; z = y - 42; if (z == 0) { crash(); }

z == 0 ? y = 42


Improving AFL

• New kind of hybrid approach

• discovering paths with AFL/QBDI

• solve unsatisfied path constraints with Triton

• Inspired by Shellphish’s Driller

• used in 2016 DARPA's Cyber Grand Challenge

�27


Scalability

�28

• Scalability is a major challenge

• path explosion (both in AFL and symbolic execution)

• amount of generated data

• Machine learning is essential to vulnerability research

• it is making it more efficient today

• it will make it more scalable tomorrow

cisco and ecole polytechnique · automatizing vulnerability research to better face new software...

Documents